Fast SDXL Image Generator Service

Description

Deploy a text-to-image pipeline as a service, using 'sdxl-turbo' by default.

Example Usage

from covalent_blueprints_ai import sdxl_basic

sdxl_blueprint = sdxl_basic()
sdxl_client = sdxl_blueprint.run()

prompt = "A beautiful sunset over the ocean."
num_inference_steps = 1

# Generate an image based on a prompt.
img_str = sdxl_client.generate(
    prompt=prompt,
    num_inference_steps=num_inference_steps,
)

# Display the image.
import base64
import io
from PIL import Image

buffer = io.BytesIO(base64.b64decode(img_str))
Image.open(buffer)

# Tear down the deployment.
sdxl_client.teardown()

Executors

text_to_image_service

GPUs:

 1

CPUs:

 25

GPU Type:

 l40

Time Limit:

 10800

Memory:

 57344

Env:

 sdxl-basic@blueprints

Environment

Pip Packages Added

torch

transformers[sentencepiece]

accelerate

diffusers

covalent-cloud

0.71.0rc0

covalent-blueprints

0.1.0

Arguments

model_name

stabilityai/sdxl-turbo

The name of the model to deploy. Defaults to "stabilityai/sdxl-turbo".

torch_dtype

float16

PyTorch data type (as string) for model parameters. Defaults to "float16".

variant

fp16

Model variant. Defaults to "fp16".

use_saved_model

true

Load the saved model from the cloud volume, if available. Defaults to True.

save_model_to_volume

false

Save the pretrained model to the cloud volume, overwriting if a copy already exists. Defaults to False.